Meet-up 2019 | Doctorants & IndustrieDEEP LEARNING
APPLIED TO ENERGY CONSUMPTIONAicha DRIDI1
Hossam AFIFI1, Jordi BADOSA 2
1 Samovar, Telecom SudParis, Institut Polytechnique de Paris 2 TREND-X, LMD, École polytechnique, 91128 Palaiseau, France
Institut Polytechnique de Paris
Aicha DRIDI- [email protected]
AbstractArtificial intelligence has initiated new opportunities for the development of various areas. The energy management
systems with the obligation of the best use of energies are with the first who should benefit from these technologies.In this paper, we propose a dual deep neural network architecture. We are trying to highlight the potential of Deep
neural networks typically Long Short Term Memory(LSTM) on making accurate time-series predictions we are also usingcategorization methods to classify time series.
1 CONTEXT
Climate issues are in the heart of vari-ous fields of researches. In this theme, thedemand for electricity and renewable ener-gies (solar, wind turbine, hydro, geothermal,...) plays a leading role. Because of theuncertainty of renewable energy production,we need to make accurate management be-tween the generated energy and consumed.=> The efficient management of these re-sources relies on forecasting the habits of en-ergy consumption.
Figure 1: Energy consumption and production for oneday
2 OBJECTIVES
Considering a Micro-grid composed of pho-tovoltaic panels, wind turbines, and otherrenewable energy sources, various buildings( energy consumers), storage capacities andaccess to the utility grid.Since energy consumption depends on sev-
eral factors such as user’s habit, external tem-perature, occupancy and so on, likewise forthe production of renewable energy becauseof their irregular and uncertain nature weneed to use advancedmethods for predictingthe future energy consumption and produc-tion to meet the adequate balance betweenthem.Our chief objective is to elaborate a smart
Energy Management System(EMS) that will
consider the different actors of a micro-gridand manage it in an efficient way.
Figure 2: Micro-grid Architecture
3 PROPOSED METHOD
In the first step, since there are manybuildings and diverse kinds of energy con-sumptions for each one. We need to doclassification. After making comparisons ourchoicewent toMulti-Layer Perceptron (MLP).
At a second step, we used Long Short TermMemory to do Time Series Predictions.
Figure 3: Comparaison SVM CNN and MLP
4 CASE STUDY
We are treating energy consumption data collected from two buildings. The first one is The DrahiXNovation Center located at Ecole Polytechnique, and the second one is The student Dorm "Maisel"located at Evry.
(a) Building with 7 different Zone (b) student dorm
Figure 4: Working Environment
5 RESULTS
Using MLP to classify the various kind of energy consumption gave us an accuracy of 0.96. Westarted by comparing Long Short TermMemory(LSTM) and Gated Recurrent Unit (GRU) to determinewhich one gave better results on the baseis of Mean Square Error (MSE), Root MSE and R2.
Table 1: Efficiency of deep prediction algorithms
KPIApproach GRU LSTM
MSE 0,007 0.006RMSE 0.076 0.063
R2 0.911 0.918
Results show that LSTM outperform GRU. In the next figures we will show predictions of the threekinds of energy consumption for the first zone.
(a) Critic Energy Consumption (b) Programmale Energy Consumption (c) Confort Energy Consumption
Figure 5: Forecasting different kinds of Energy consumption
6 FUTURE WORK
After starting by categorizing the kind of energy consumption and the type of building, we traineda Deep Neural Network on historical data collected from two distinct buildings to forecast one day-ahead energy consumption and photovoltaic panel production.
Figure 6: Classification and Prediction
The next step is to develop themodel using either optimizationmethods or Reinforcement learning.
References[1] Anastasia Borovykh, Sander Bohte, and Cornelis W Oosterlee. Conditional time series forecasting with
convolutional neural networks. arXiv preprint arXiv:1703.04691, 2017.[2] Kyunghyun Cho, Bart Van Merriënboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Holger
Schwenk, and Yoshua Bengio. Learning phrase representations using RNN encoder-decoder for statis-tical machine translation. arXiv preprint arXiv:1406.1078, 2014.
[3] Tae-Young Kim and Sung-Bae Cho. Predicting residential energy consumption using cnn-lstm neural net-works. Energy, 2019.
[4] Zheng Zhao, Weihai Chen, Xingming Wu, Peter CY Chen, and Jingmeng Liu. LSTM network: a deep learn-ing approach for short-term traffic forecast. IET Intelligent Transport Systems, 2017.
